OpenAI unveils custom Jalapeño AI chip, analysts say Nvidia faces margin threat

OpenAI said the chip has "industry-leading speed and efficiency" and will allow users to get "faster responses, more responsive agents and more reliable access" as demand grows. It will be deployed within OpenAI's compute infrastructure by the end of the year, and OpenAI said it was already working on generations two and three. The company first announced Jalapeño in June, describing it as built from the ground up for current and future large language models.

Nvidia's share price has surged amid the data center buildout and huge demand for its chips in model training and inference. But hyperscalers and AI companies are increasingly developing custom silicon. Google, AWS and Meta are also building their own AI chips.

Adrien Sanchez, technology analyst at Yole Group, told CNBC that the Jalapeño chip shows a "hyperscaler-designed chip can now match or beat Nvidia's Blackwell-class GPUs on inference efficiency." While Nvidia still owns the "vast majority" of AI compute and benefits from the software ecosystem lock-in of CUDA, Sanchez said the new chip is a "threat to Nvidia's inference margins, which is the field growing the most at the moment." Nvidia has been approached for comment.

Alexander Harrowell, senior principal analyst at Omdia, called Jalapeño an "impressive achievement, most of all in terms of efficiency." In a large-scale deployment, he told CNBC, "this would save power, cooling, and power distribution infrastructure, and contribute a lot to their unit economics." TrendForce analyst Fion Chiu said the custom chip could reduce OpenAI's reliance on Nvidia over time for inference workloads, but Nvidia GPUs would remain important for compute-intensive work such as large-scale model training because of their programmability, performance, software ecosystem, and broad workload handling.

Research firm SemiAnalysis said it visited OpenAI's labs to benchmark Jalapeño and found it beat Nvidia's Blackwell on performance per watt in nearly all tested scenarios. But the firm called the comparison "somewhat incomplete and unfair" because Jalapeño uses newer HBM4 memory; it said Nvidia's Rubin platform, which also uses HBM4, is a better comparison. It also noted that Vera Rubin systems are starting to ship to customers now, while OpenAI still has only engineering samples of Jalapeño.

OpenAI has been one of the largest consumers of Nvidia GPUs, and Sanchez said Jalapeño raises the stakes for Nvidia's largest customer relationship specifically. OpenAI is among several companies pursuing custom application-specific integrated circuits, or ASICs. In April, Google unveiled tensor processing units for AI training and inference, Meta agreed to deploy 1 gigawatt of custom AI chips using Broadcom technology, and Anthropic committed more than $100 billion over 10 years to AWS, including Amazon's custom Trainium chips.

Harrowell said Omdia expects custom ASICs such as Jalapeño to exceed GPUs in volume by 2028, though revenue will take longer because GPUs are considerably more expensive. He called this "the biggest competitive threat to Nvidia," since roughly half of capital expenditure on AI infrastructure comes from hyperscale cloud providers that either have a custom chip program or could reasonably have one. Startups including Cerebras, SambaNova, D-Matrix, Etched and Fractile are also developing AI chips.